Sensor-Driven Surrogate Modeling and Control of Nonlinear Dynamical Systems Using FAE-CAE-LSTM and Deep Reinforcement...

Area of Science:

• Cyber-physical systems
• Nonlinear dynamics
• Computational fluid dynamics

Background:

• Real-time control of nonlinear partial differential equations (PDEs) is challenging due to sparse data and high dimensionality.
• Deep reinforcement learning (DRL) shows potential but requires computationally intensive training on full-order simulations.

Purpose of the Study:

• To develop a computationally efficient, sensor-driven framework for DRL-based control of nonlinear dynamical systems.
• To enable real-time control using non-intrusive reduced-order modeling (NIROM).

Main Methods:

• Introduced FAE-CAE-LSTM: a framework combining autoencoders and LSTM for state compression and temporal evolution.
• Trained DRL agents in a reduced latent space using sensor-like spatiotemporal measurements.
• Employed a CNN-MLP reward estimator for data-driven feedback without needing governing equations.

Main Results:

• Demonstrated accurate state reconstruction and robust control on benchmark systems (e.g., Burgers' equation, flow past cylinders).
• Achieved significant computational speedup compared to traditional simulation-based training.
• Validated the effectiveness of the FAE-CAE-LSTM surrogate for scalable DRL control.

Conclusions:

• The FAE-CAE-LSTM framework effectively enables real-time, sensor-informed, and scalable DRL-based control.
• This approach overcomes limitations of sparse data and high dimensionality in cyber-physical systems.
• The method offers a significant advancement for controlling complex nonlinear dynamical systems.